DGX Spark vs NVIDIA RTX Spark N1X laptops: specs, tok/s, GPU memory split, prices and software

The premise holds: the RTX Spark N1X is NVIDIA's laptop sibling of the DGX Spark's GB10, with a 20-core Grace CPU, a 6,144-core Blackwell GPU and up to 128 GB of LPDDR5X at 300 GB/s, and it ships in Windows 11 laptops that NVIDIA launched on 16 October 2026. In The Register's same-test run an HP OmniBook Ultra 16 decoded Qwen3.6-35B-A3B at 67.64 tok/s against 63.77 on a DGX Spark, while the Spark reached the first token faster and finished Artificial Analysis's agent tasks 3 to 8% sooner. The 128 GB models cost about $5,900 to $6,230 against $6,950 for the Spark. What does not carry over is the software: no official Linux, a fixed Windows GPU memory split, and no DGX OS, NGC containers or Linux vLLM. Checked 7 November 2026.

All speeds are third-party reports; GPUwerk has not tested an N1X laptop. The Register ran both machines with the same llama.cpp test on 16 October 2026, the laptop on Windows and the Spark on Ubuntu 24.04.5.

What is the RTX Spark N1X, and how close is it to GB10?

Close on paper: the same CPU layout, the same GPU core count and slightly faster memory, at a third to a half of the power. NVIDIA's RTX Spark page calls the platform the RTX Spark Superchip and lists the part as the RTX Spark N1X; its porting guide gives the memory as a 256-bit, fully coherent interface at 300 GB/s.

NVIDIA RTX Spark page, porting guide and DGX Spark hardware page, read 7 November 2026; the 18-core figures are from StorageReview's Lenovo Pro 9n 15 review. NVIDIA's page also lists a desktop N1X with an 18-core CPU, a 5,120-core GPU and up to 64 GB, and gives 140 W in a second, unlabelled table. The GPU identification page lists separate PCIe device IDs for the laptop and desktop 6,144-core parts. The Spark's 6,144 CUDA cores are from the specs page.

The Register's review explains the bandwidth gap: GB10 supports 9,400 MT/s memory, but the DGX Spark ships at 8,533 MT/s, and the N1X laptops run the faster memory. Measured, it is less than 300 GB/s from the CPU side: StorageReview's STREAM Triad gave 122.1 GB/s on the Surface Laptop Ultra and 120.3 GB/s on the 18-core Lenovo. How that compares with GB10 measured the same way is the subject of the sibling guide on measured DGX Spark bandwidth.

Which RTX Spark laptops are on sale, and what do they cost?

Six makers are listed by NVIDIA; I could confirm prices for three. Prices are US dollars before sales tax unless marked.

Microsoft's FAQ lists 24, 32, 48, 64 and 128 GB configurations; the 128 GB price is from the two reviews, not from Microsoft's page. Geizhals blocked my own access on 7 November 2026, so the EUR figure is the one recorded in the linked article that day.

At 128 GB the laptops and the Spark cost about the same. The smaller configurations are where the laptops are cheap, but a 24 or 32 GB N1X gives its GPU less memory than an RTX 5090 laptop's 24 GB of VRAM (StorageReview's 32 GB Lenovo was split 12 GB for the GPU) at a third of its 896 GB/s, as the RTX 5090 laptop article shows.

How fast is an N1X laptop against a DGX Spark?

About 6% faster at generating tokens and slower to the first token, so end to end the two finish agent work within a few percent. These are the published runs I found.

Artificial Analysis notes that each system used its own recipe (runtime, speculative decoding, weight format), so its comparison is not like for like, and that its laptop was a prototype. The Register names llama.cpp v0.6.0 for its single-batch test and does not say whether its figures are medians or means.

The decode lead fits the memory: 300 against 273 GB/s is 10% more bandwidth, and the measured leads are 6.1% and 6.3% (my division). The first-token deficit fits the power: prompt processing is compute-bound, and the laptop part runs at 45 to 80 W against GB10's 140 W. In the agent run the Spark was about 1.55 times faster to the first token, which is the likely reason it wins Artificial Analysis's end-to-end timings despite decoding a little slower. For an RTX 5090 desktop the same benchmark gave 2.0 minutes on Qwen3.6 35B A3B, which is what a 1,792 GB/s card does on a model that fits in 32 GB.

How much of the 128 GB can the laptop's GPU use?

62.5 GB by default on the reviewed 128 GB Surface, up to 96 GB if you change the split; the Spark allocates its memory on demand. StorageReview's unit arrived split 64/64 between system and GPU and reported 62.5 GB usable by the GPU. NVIDIA's unified memory chapter describes three regions on Windows: a dedicated GPU carveout, shared memory that both processors can use, and CPU-only memory. Its CUDA chapter says cudaMalloc fills the dedicated segment first and then spills into the shared one, whose smaller pages can make access slower, and advises against cudaMallocManaged. The Register adds that Windows does not reassign GPU memory dynamically the way Linux does on the Spark.

# does the model fit the GPU carveout? weights + cache at 32k + 1.5 GB runtime (my assumption)
gpt-oss-120b MXFP4     63.39 + 1.21 + 1.5 = 66.1 GB  -> over the default 62.5 GB carveout; fits a 96 GB split
Qwen3.5-122B-A10B Q4   77.03 + 0.81 + 1.5 = 79.3 GB  -> needs the 96 GB split
Qwen3.6-35B-A3B Q4     22.36 + 0.67 + 1.5 = 24.5 GB  -> fits either split
DGX Spark: about 110 GB usable for all three, no split to set

File sizes from Hugging Face for gpt-oss-120b and Qwen3.5-122B-A10B (UD-Q4_K_XL), read 7 November 2026; cache sizes from each model's config.json at FP16. The Register ran Qwen3.5-122B-A10B on its 50:50-split OmniBook at 26.23 tok/s, so llama.cpp spilled into shared memory or the split was raised; the article does not say which.

A 96 GB GPU split leaves 32 GB for Windows and everything else, which is workable on a laptop and not generous. On the Spark the same gpt-oss-120b file simply loads, with room left for a long context.

Does the CUDA and Linux stack match the DGX Spark?

CUDA does; the rest of the Linux stack does not. NVIDIA's RTX Spark page says CUDA runs natively and lists Windows 11 as the supported OS. CUDA Toolkit 13.4 added native Windows on Arm builds on 9 September 2026, according to ComputingForGeeks, which also reports that the N1X GPU identifies as compute capability 12.1, the same as GB10; I did not find that figure on NVIDIA's own pages, so treat it as reported. The RTX Spark developer preview (July 2026) shipped a Windows 11 Arm64 driver and a CUDA 13.4 preview, with known issues that include possible instability during PyTorch build workflows and slow transfers from pageable memory.

• Linux: not announced by NVIDIA, per ComputingForGeeks on 3 October 2026. The community awesome-rtx-spark list (reviewed 17 October 2026) describes native Linux as community bring-up only, with a patched kernel tested on the ASUS ProArt P14 and Dell XPS 16, and a firmware default that reserves about 62.5 GiB for the GPU.
• What works on the laptop: native Arm64 llama.cpp and Ollama, which StorageReview used, and NVIDIA's llama.cpp recipes for RTX Spark, which Artificial Analysis used. x64 Windows apps run through Microsoft's Prism emulation.
• What is Spark-only: DGX OS, NGC containers, the Linux vLLM and SGLang builds, NVIDIA's DGX Spark playbooks and the ConnectX-7 two-node cluster described in the two-node guide. The Register notes that PyTorch support is better on Linux.

So code you write against CUDA can move between the two machines, but a serving setup built on the vLLM container on the Spark does not move to the laptop. If you mainly use LM Studio or Ollama, the difference shrinks; the LM Studio vs Ollama comparison covers those two tools.

How does a laptop N1X hold up under sustained load?

Hot, and in one reviewed unit not stable on an all-core CPU test. StorageReview's Surface Laptop Ultra held its CPU at 97 to 98 °C through a 31-minute sustained loop without losing speed, but an all-core OpenBLAS solve cut power to the second review unit in three of four attempts, logged as Kernel-Power 41 with no thermal-trip event; the first unit had failed the same way and was replaced. StorageReview reported the Surface fault to the vendor. The 18-core Lenovo Pro 9n peaked at 93 °C and averaged 87 °C in its sustained loop. These were CPU tests, not LLM runs.

GB10 desktops have had their own power-off reports under long GPU loads in 2026, compared by maker in the sibling guide on GB10 sustained performance by OEM. Neither platform is immune; the laptop has a third to a half of GB10's power budget and a smaller cooler.

Verdict: laptop or Spark?

• You need one portable machine for Windows apps and local models up to about 60 GB: an RTX Spark laptop. Decode speed matches a Spark, and the 128 GB configuration costs a little less.
• You serve a model to other people, use vLLM, SGLang or NGC containers, or want two units clustered: the DGX Spark. The Linux stack is the product, and the laptop does not run it.
• Models of 60 to 110 GB: the Spark, unless you are happy to raise the laptop's GPU split to 96 GB and live with 32 GB for Windows.
• Long prompts and RAG: the Spark, at 1.55 times faster to the first token in The Register's agent run.

For other 128 GB alternatives, see DGX Spark vs Mac Studio and DGX Spark vs Strix Halo. If the Linux side is what you need, a rented Spark from GPUwerk runs the same GB10 by the minute, which is a cheap way to check your model before spending $6,000 on either.

FAQ

Sources

Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.

• NVIDIA RTX Spark, NVIDIA
• System Overview, NVIDIA Windows on Arm Porting Guide, NVIDIA
• Unified Memory Architecture, NVIDIA Windows on Arm Porting Guide, NVIDIA
• Usage with CUDA APIs, NVIDIA Windows on Arm Porting Guide, NVIDIA
• RTX Spark GPU Identification, NVIDIA Windows on Arm Porting Guide, NVIDIA
• DGX Spark hardware overview, NVIDIA
• Benchmarked: Nvidia's RTX Spark notebook delivers serious local AI compute, at a price, The Register, 2026-10-16
• First AA-AgentPerf-Local results for an NVIDIA RTX Spark N1X laptop, Artificial Analysis, 2026-10-15
• Microsoft Surface Laptop Ultra for Business Review: RTX Spark N1X, StorageReview, 2026-10-15
• Lenovo Pro 9n 15 Review: 18-Core RTX Spark, StorageReview, 2026-10-16
• Surface Laptop Ultra, Microsoft
• RTX Spark Developer Preview, NVIDIA Developer Forums, 2026-07-16
• NVIDIA RTX Spark Laptops: N1X Specs, CUDA and Linux, ComputingForGeeks, 2026-10-03
• awesome-rtx-spark, Th0rgal on GitHub, 2026-10-17
• gpt-oss-120b-GGUF files, ggml-org on Hugging Face
• Qwen3.5-122B-A10B-GGUF files, Unsloth on Hugging Face